1

I have been training an Object detection model on IBM PowerAI Vision. When i train the model i see the following marked losses. i want to know in deep learning terms what are these losses. Since there are many kinds of losses used for object detection. My initial research pointed me towards

  • Cross entropy = CLS loss
  • IoU loss = BBox loss

But i am not sure, documentation doesn't say much about it. enter image description here

Qaisar Rajput
  • 751
  • 1
  • 8
  • 21

1 Answers1

2
  • CLS = Classification, Localization, Segmentation. Stands for combined error measurement in R-CNN models. Shows how accurate trained model can split original image into smaller regions, select (localize) most interesting of them and classify spotted objects (if there is any).
  • BBox = bounding box. Measures how precisely trained model can put bbox coordinates on recognized object, compared to test subset.

Both CLS and BBox measured several times while training and the last one specifies resulting error levels.

For details & maths see https://lilianweng.github.io/lil-log/2017/12/31/object-recognition-for-dummies-part-3.html

Pavel
  • 46
  • 5